Justification

3. A statistical analysis to select the classifier that provides the best performance during the training stage of the BCI.

4. A BCI based on the combination of the proposed methods, in order to preserve the information of interest on EEG signals, providing a better calibration stage to improve the pattern recognition.

5. An evaluation on the information preservation of the proposed adaptive filter throughout an SSVEP database.

6. An experimental protocol to study motor patterns related to gait planning/stopping.

7. An evaluation of several classifiers, such as Linear Discriminant Analysis (LDA), Reg- ularized Discriminant Analysis (RDA) and Support Vector Machine (SVM) with kernel lineal, coming to obtain the best BCI to recognize gait planning/stopping.

1.5

Justification

Lower-limbs robotic exoskeletons may be used to assist people with neural injury, such as spinal cord injury, stroke, among others (PONS, 2008). Additionally, a BCI can be used to convey con- trol commands from the brain for a robotic exoskeleton to reproduce more natural movements after a voluntary intention, aiming to produce neuroplastic changes in the cortical pathways on these subjects (JIANG et al., 2015; HORTAL et al., 2016; VELU; SA, 2013; SBURLEA; MON- TESANO; MINGUEZ, 2015). For this purpose, patterns related to pre-movement actions or motor planning (from -2 s before the movement onset), such as ERD/ERS (PFURTSCHELLER; SILVA, 1999) and MRCP (SHIBASAKI; HALLETT, 2006) potentials, may be used to recog- nize motor intention with very short delay, so that the resulting control can be perceived as a closed loop (XU et al., 2014).

Events related to motor intention, such as ERD/ERS, can be blurred by the poor SNR of EEG, due to several sources of interferences and artifacts, such as electrode and eyes move- ments, eyes blink, myoelectric and cardiac activities, non-µ-rhythm of EEG components as

visual alpha rhythm, among others (MCFARLAND et al., 1997; KILICARSLAN; GROSS- MAN; CONTRERAS-VIDAL, 2016). Moreover, the motor anticipation recognition may be affected by process of high uncertainty, such as pattern labeling, as these events are located before any muscular activation (from -2.0 to 0s), where it is not possible to use goniometer and footswitch, in order to select correctly these patterns. Therefore, BCIs based on unsupervised methods may be suitable to reduce effects of these issues.

Several BCIs have been proposed to recognize gait planning (SBURLEA; MONTESANO; MINGUEZ, 2015; SBURLEA et al., 2015; SBURLEA; MONTESANO; MINGUEZ, 2017; DELISLE-RODRIGUEZ et al., 2017), onset (JIANG et al., 2015; HORTAL et al., 2016; VELU; SA, 2013), and stopping (HORTAL et al., 2016; DELISLE-RODRIGUEZ et al., 2017). BCIs for gait onset recognition have been more explored, for example, an MRCP template extracted from several trials have been used in a BCI based on the Matched filter to detect gait onset (JIANG et al., 2015). Here, ICA and Laplacian filters were also applied in the pre-processing stage to improve the SNR of EEG signals, improving the BCI performance with true positive rate (TPR) and false positive rate (FPR) of 76.9 ± 8.97% and 2.93 ± 1.09 per minute, respectively. MRCPs occur with small amplitudes (typically lies between 5 and 30 µV) at low frequency (from 0.1 to 4 Hz) (SHIBASAKI; HALLETT, 2006; JIANG et al., 2015), therefore, it may be masked by potentials of higher frequency band. Thus, BCIs based on MRCPs need several trials in the calibration stage to obtain an averaged MRCP, in order to remove the background noise of each trial (SBURLEA; MONTESANO; MINGUEZ, 2015).

In addition, a BCI based on the MRCP phase representation was proposed to decrease the time-consuming in the calibration stage, enhancing the gait intention recognition (SBURLEA; MONTESANO; MINGUEZ, 2017). As a result, a comparison of three BCIs based on MRCP amplitude, MRCP instantaneous phase, and a combination of both representation was carried out, where the combination of both MRCP features showed the best performance (accuracy of 66.5% on healthy subjects, and 63.3% on stroke patients).

Other BCIs that compute features on mu and beta frequency bands using the power spectral density method have been proposed to recognize gait intention (DO et al., 2013; HORTAL et al., 2016), with accuracy (ACC) of 68.6% and 86.3%, respectively. Both researches were con-

1.5. Justification 13

ducted on subjects with spinal cord injury, but three classes (resting, starting, and stopping) were considered in (HORTAL et al., 2016).

Although patterns related to motor planning may improve the closed-loop during the user- exoskeleton interactions, few BCIs based on motor planning have been proposed (SBURLEA; MONTESANO; MINGUEZ, 2015; SBURLEA et al., 2015). For example, a combination of both MRCP and ERD/ERS was evaluated on nine post-stroke patients, in order to recog- nize pre-movement states (ACC of 64%, in a range from 18 to 85.2%) by a BCI during walking (SBURLEA et al., 2015). It is worth noting that patterns related to motor plan- ning on gamma-ERS band were used (PFURTSCHELLER et al., 2003; SEEBER et al., 2015; PFURTSCHELLER; SILVA, 1999). Although, this band has been little explored to recognize gait intention, researches have recently demonstrated that high and low gamma-ERS bands may be used for gait analysis (SEEBER et al., 2015; COSTA et al., 2016).

Additionally, several works have been conducted during walking to recognize action types and directions (VELU; SA, 2013), as well as to detect unexpected obstacles (SALAZAR-VARAS et al., 2015). For these objectives, coefficients from 9 levels of discrete wavelet decompositions using Daubechies Wavelet 4 as a mother wave were computed (VELU; SA, 2013), where coeffi- cients obtained at low frequency bands were highly weighted. This evidences an important role of low frequency components, also when using slow cortical potentials (SCP) or MRCPs, to recognize motor intention, which agrees with (JIANG et al., 2015; SBURLEA; MONTESANO; MINGUEZ, 2015; LEW et al., 2012; SHIBASAKI; HALLETT, 2006).

Regarding BCIs specially designed for gait intention recognition, several authors have com- monly used classifiers, such as Linear Discriminant Analysis (LDA) (JIANG et al., 2015), Regularized Linear Discriminant Analysis (RDA) (VELU; SA, 2013) and Support Vector Ma- chine (SVM) (HORTAL et al., 2016). Furthermore, pre-processing methods, such as ICA, and Laplacian filters have been adopted to improve the SNR of EEG signals (HORTAL et al., 2016; JIANG et al., 2015; VELU; SA, 2013). Moreover, since EEG signals are highly variable be- tween subjects and recording sessions, typically BCIs are calibrated at the beginning of each session. This calibration stage should be short, especially when patients that suffering loco- motor disabilities cannot maintain a bipedal position for a long time. To improve this issue,

a BCI to execute continuous recognition of gait onsets without session-to-session recalibration was proposed (SBURLEA; MONTESANO; MINGUEZ, 2015). On the other hand, BCIs based on unsupervised learning to recognize gait planning/stops have been little explored.

BCIs with acceptable speed are suitable to obtain closed-loop control that produce more nat- ural movements after a voluntary intention, which provide a neuroplastic change in the corti- cal pathways on people with neural injury. Previous works have demonstrated that the BCI speed and accuracy can be affected by the SNR achieved through spatial and temporal fil- tering methods (MCFARLAND et al., 1997). In the literature, few adaptive methods have been reported for BCIs, in order to improve the motor patterns recognition during dynamic movements, as walking. Nowadays, almost all BCIs developed on system for gait rehabilita- tion/assistance, are based on ICA and spatial filters based on Euclidean distance (Laplacian, LAR and WAR) (JIANG et al., 2015; HORTAL et al., 2016; VELU; SA, 2013). ICA can be also applied for real time applications but the performance of real-time data over long pe- riod of time may be affected, as it is practically impossible to obtain a mixing matrix from a training database that contains all possible physiological and non-physiological artifacts (KILI- CARSLAN; GROSSMAN; CONTRERAS-VIDAL, 2016). Furthermore, the operation of Lapla- cian, LAR and WAR filters does not depend on underlying data, which may contribute to add on the electrode of interest undesirable artifacts.

Other adaptive filters have been proposed to remove artifacts on EEG signals (KILICARSLAN; GROSSMAN; CONTRERAS-VIDAL, 2016; KHATUN; MAHAJAN; MORSHED, 2016), but they have not been applied in BCIs for gait intention recognition. For example, principal com- ponent analysis (PCA) have been used to remove artifacts in offline analysis (JUNG et al., 1998b). These BCIs based on ICA and PCA have demonstrated a good performance in motor intention recognition. However, additional stages, such as artifact annotation by trained spe- cialist and new classifiers must be included to identify possible patterns correlated to artifacts throughout EEG.

In particular (MULLEN et al., 2013), a method based on PCA, denoted as artifact sub- space reconstruction, was proposed to remove artifacts on a limited time window of EEG, which yields promising results for real-time applications. Moreover, adaptive noise cancelling

1.5. Justification 15

(ANC) methods have been developed for signal extraction from noise corrupted measurements, in which several adaptive estimators as recursive least squares and Kalman filter have been used (VASEGHI, 2008; SWEENEY et al., 2012; SWEENEY; MCLOONE; WARD, 2013; GUERRERO-MOSQUERA; NAVIA-VÁZQUEZ, 2012). Many of these methods do not adopt the non-stationary behavior of EEG signals, which is considered as time-varying frequency char- acteristics in general. This condition may not exist in some cases, however, it is likely to exist in EEG recordings over duration of an experimental session, and across sessions. These fluc- tuations create time-varying characteristics over the course of an experimental session, which require adaptive and robust filters to reject these abnormal events.

Additionally, several studies have used standard electrode locations close to the ocular source, such as frontal electrodes FP1, FP2, FT9, and FT10 to provide a reference input with EOG components for methods as ANC to remove artifacts (PUTHUSSERYPADY; RATNARAJAH,

2005). Other authors have used H∞ to obtain an adaptive filter using EOG components ac-

quired around the eyes as reference input, which considers full or partial superposition of arti- facts, such as eye blink, eye movement, and signal drift throughout EEG (KILICARSLAN;

GROSSMAN; CONTRERAS-VIDAL, 2016). Furthermore, another method has been pro-

posed for a single channel scenery, using the wavelet transform decomposition (KHATUN; MAHAJAN; MORSHED, 2016). These aforementioned methods have demonstrated a good performance to remove EOG artifacts on EEG, preserving the neural information on the re- constructed segments.

In summary, one of the most important challenges for BCIs to command robotic exoskeletons is to provide a closed-loop control with more natural movements, as well as reduce time-consuming related to session-by-session re-calibration (or training). Thus, methods of low computational cost, robust to artifacts and uncertainty due to pattern labeling must be explored, in order to obtain a better BCI performance. In spite of the large number of stroke incidents affecting the locomotion, few attention has been devoted to study the pre-movement recognition during walking.